Research On Recommendation Algorithm Based On Privacy Preservation

Recommendation system is a successful technology that has been implemented in E-commerce recommender system.It can effectively alleviate information overload problem brought by the rapid development of Internet,and it can dig users' potential demand from massive data according to users' behavior or preference.Collaborative filtering(CF)algorithm is the most widely used algorithm.It is based on the assumption that a user will prefer items that similar users prefer.However,collaborative filtering algorithm is also vulnerable to malicious users,such as shilling attack and kNN attack.In shilling attack,attacker will create a certain number of fakers whose profile are similar to normal users to disrupt the prediction accuracy of recommendation algorithm,such as promote or demote some items' ratings;In kNN attack,attacker will construct some fakers who have similar ratings as the target user on some items to get target user's sensitive information.No matter which attack occurs,users' benefit will be damaged and then users will lose the trust for recommendation system.Therefore,privacy preservation of recommendation algorithm has become a research hotspot.Aiming at shilling attack and kNN attack,this paper conducts further research and proposes solutions.The main research achievement as follows:First,this paper researches on the current algorithms against shilling attack according to the implementation and attack feature of shilling attack.Current solutions mainly focus on attack detection methods and robust CF algorithms.To overcome the flaws of higher false positive and unassured prediction accuracy,this paper proposes a soft-decision method which gets user's suspicion probability by applying SVM and generates partitions of variable sizes.This method chooses neighbors that are similar to target user with the given security metric.This method reduces false positive rate through marking suspicious fakers instead of deleting them directly such that misclassified normal users can still contribute to the similarity calculation.Experimental analysis shows that our approach ensures excellent prediction accuracy against shilling attack.Second,this paper researches on the current algorithms against kNN attack according to the implementation and attack feature of kNN attack.Current solutions mainly focus on cryptographic methods,obfuscation methods and perturbation methods.In this paper,to overcome the flaws of unnecessary computational cost,lower data quality and the difficulty to calibrate the magnitude of noise,we mainly study the privacy preservation of collaborative filtering algorithm based on k-anonymity method,and propose a novel algorithm based on the data characteristics of recommendation algorithm.This method improves microaggregation algorithm based on importance partitioning to increase homogeneity among records in each group which can retain better data quality,and proposes(p,l)-diversity and(p,l,?)-diversity model where p is attacker's prior knowledge about users' ratings,l and(l,?)are the diversity among users in each group to improve the level of privacy preservation.Experimental analysis shows that our approach ensures a higher level of privacy preservation based on lower information loss.